Abstract
3D plant phenotyping has emerged as a crucial enabler of modern precision agri culture, offering detailed structural and functional insights into plant development for applications such as stress detection, yield estimation, and genotype evaluation. While traditional 2D imaging techniques lack the capacity to represent complex plant geometries, 3D point cloud analysis provides richer spatial information necessary for high resolution phenotypic studies. Despite significant progress in 3D plant phenotyping, several challenges remain. Pub licly available annotated 3D plant datasets are scarce and often constrained in terms of species diversity and dataset scale, limiting the capacity to develop generalisable deep learning models. Furthermore, the structural complexity of plants, coupled with intra-class variation, particularly in organs such as leaves and spikes (ears) pose ongoing difficulties for accurate part segmentation. Weintroduce Wheat3D PartNet, alarge-scale annotated 3D point cloud dataset of wheat plants, comprising 1,303 models across three cultivars (Paragon, Gladius, and Apogee) grown under varying environmental conditions. Each sample is manually annotated into biologically meaningful parts, namely spikes (ears) and leaves/stems (non-ears), to support plant part segmentation and trait estimation tasks. Unlike existing datasets, Wheat3D PartNet captures cultivar-specific variation with biolo gically informed labels, enabling high resolution part based phenotyping in complex crop structures. To handle the prevalent issue of intra-class imbalance, we first analysed the distribu tion of annotated points across plant parts. In Wheat3D PartNet, the proportion of ear points varies considerably across cultivars. In Paragon, ear points constitute 0 to 25% of the total point, Gladius shows a moderate range of approximately 12%–37%, whereas Apogee presents a relatively balanced distribution, with 22%–50% of points belonging to the ear class. This uneven representation limits the model’s ability to learn discriminative features for the underrepresented spike part. We propose a feature-based sampling and loss weighting strategy that lever ages biologically relevant traits such as ear count and ear ratio. This approach improves model sensitivity to underrepresented parts, such as wheat spikes (i.e., ears). On average, this method improved ear mIoU by 3%–4% across cultivars, demonstrating its effectiveness in mitigating part level imbalance. Furthermore, we introduce a Monte Carlo Dropout-based uncertainty-driven samp ling strategy that identifies ambiguous and highly uncertain training samples. These samples are selectively reintroduced during training, enabling the model to focus on structurally challenging or underrepresented regions, thereby improving ro bustness and generalisation.This method led to an average increase of 4.5% in overall mIoU compared to the baseline. Building upon these findings, we propose an uncertainty-aware and feature based loss function that combines epistemic uncertainty with biologically informed trait weighting, specifically emphasising wheat spikes (i.e., ears). When applied to baseline models, this hybrid loss function resulted in ear mIoU improvements of up to 15.3%, highlighting its effectiveness in learning discriminative features for structurally complex plant organs. Finally, we present WheatAdaptNet, a novel segmentation network that integrates Multi-Head Adaptive Sampling (MHAS) and a Geometric Attention (GA) block. MHAS learns head-specific spatial offsets to form adaptive neighbourhoods, captur ing diverse local geometric patterns, while the single GA block emphasises salient structural features. This design reduces computational overhead compared to multi stage attention, improves training stability, and enhances cultivar-adaptive segment ation performance. Experimental results across all three wheat cultivars demonstrate consistent improvements in ear mIoU and overall segmentation accuracy compared to state-of-the-art models, with WheatAdaptNet outperforming GAPNet and other baselines by up to 4% to 7% in ear mIoU.This thesis contributes a scalable dataset, introduces novel imbalance handling tech niques, and develops an efficient adaptive attention-based segmentation model, lay ing a foundation for robust 3D plant segmentation methods applicable in real-world phenotyping and precision agriculture scenarios
| Date of Award | 2026 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | YONGHUAI LIU (Director of Studies), HUAIZHONG ZHANG (Supervisor) & John H Doonan (Supervisor) |